Uncertainty-driven Fusion for Conflictive Multiview Data: Beyond View Alignment Assumptions

Abstract

Multiview learning aims to integrate diverse feature representations to achieve a comprehen- sive understanding of data. Traditional approaches often assume strict alignment across views, making them ill-suited for real-world scenarios where low-quality conflictive instances, i.e. in- stances with conflicting information across views are prevalent. Existing methods largely focus on eliminating conflicting instances by discarding them or substituting conflicting views, over- looking the need for practical decision making in such cases. Furthermore, while the recently proposed Reliable Conflictive Multiview Learning (RCML) framework introduces the idea of attaching reliabilities to decision outcomes, it leaves certain theoretical gaps unaddressed, es- pecially prioritization of conflictive views in fusion process in a principled manner.